The Council of the EU has now adopted its common position on the bloc's draft AI Act, as the legislation approaches final form. The updated position signals a number of important refinements to key principles in the previous text.

The EU is hoping the legislation, once agreed and ratified with the European Parliament, will provide organisations within or trading into the EU with a uniform, risk-based framework, creating a single market for AI applications. 

Key points flagged in the Council's approach include:

• amendments to the definition of an 'AI System', seeking to evolve the previous definition by distinguishing AI from simpler software systems, referring to machine learning and logic- and knowledge-based approaches as opposed to more conventional software applications;
• extension of the prohibition on use of AI for social scoring, to apply this to private as well as public organisations;
• refinements of the classes of AI systems considered as high-risk, to ensure that systems that are not likely to harm fundamental rights or create significant risks are not captured by the legislation;
• clarification of the requirements applicable to those high risk AI systems which are also capable of general purpose uses - a further implementing act will define the application of these principles to general purpose AI. Importantly this includes helpful clarifications around the degree of assurance needed by SMEs in order to demonstrate their compliance with the Act, something that the Council no doubt intends to promote innovation throughout the AI supply chain;
• predictably following the more detailed input of individual Member States through the Council, an exclusion of the areas of national security, defence and military applications from the scope of the Act;
• an increase in the transparency requirements relating to use of high risk AI systems, including obligations on public bodies to register against an EU database when deploying high risk AI; and
• measures to support innovation in the AI marketplace, including provision of regulatory sandboxes to allow new AI systems to be developed and tested in 'real-world' scenarios.

The Council will now enter the next phase of negotiations with the European Parliament, which itself needs to finalise its position on the draft Act, the hope being that agreement will be reached between the institutions in the early part of 2023 to allow for implementation later next year.

Whether the horse-trading of this final phase results in further material changes remains to be seen, but either way the shape of the Act appears to be coming into sharper focus and, with material regulatory and financial penalties attached (see previous updates from our Shoosmiths AI Working Group for details), organisations developing or procuring AI-based systems will need to ensure they are set up to demonstrate compliance.